Free Sample Lesson: How an agent works: the building blocks of Agent Script (Agentforce Specialist Cert Lab)

Free sample lesson from the Agentforce Specialist Cert Lab. The full course has 24 lessons and 240 practice questions.

Get the Cert Lab, $9.99

AI Agents · AG1 · part of the 35% section · about 25 min with practice

AG1: How an agent works: the building blocks of Agent Script

You’ll be able to: name the blocks in an Agent Script file, say what each one controls, and trace how a request moves from the agent router into a subagent.

Diagram: a customer message goes to the start_agent block (the Agent Router), which routes to a subagent; the subagent resolves its instructions into a prompt and the LLM responds. The next message starts at the router again.

Agent Script is the language for building agents in Agentforce Builder. An agent is one script file made of top-level properties called blocks. When you save a version of the agent, the script compiles into lower-level metadata that the reasoning engine uses.

Terminology note: beginning in April 2026, agent topics are called subagents, with no change in functionality. The Agentforce Specialist exam guide still says “topics”, so treat the two words as the same thing.

The blocks you need to know

Block	What it holds	Remember
`system`	General instructions and system messages	`welcome` and `error` messages are required
`config`	developer_name, agent_label, description, agent_type and more	developer_name must be unique in the org (max 80 chars)
`access`	`default_agent_user`	Required for Agentforce Service agents
`variables`	Global variables every subagent can read	Referenced as `@variables.name`
`language`	Default and additional locales	Voice needs a fixed default locale
`connection`	How the agent works with an outside channel, such as messaging escalation	Pairs with `@utils.escalate`
`start_agent`	The Agent Router: classification and routing	Every customer utterance begins here
`subagent`	description, reasoning (instructions + tools) and actions for one job	The description helps routing pick it

Inside a subagent

A subagent block has a name (snake_case, no spaces), a description that helps the agent decide when the subagent fits the user’s intent, a reasoning section and an actions section. reasoning.instructions guides the reasoning engine once it picks the subagent. reasoning.actions lists the tools the reasoning engine may use. The subagent’s actions block defines the actions themselves (inputs, outputs and target). To let the reasoning engine use one of those actions, you must also point to it from reasoning.actions.

How a request flows

All requests, including the first one, begin at the start_agent block. It typically sets initial variable values and does subagent classification. Agentforce then processes the chosen subagent’s reasoning instructions from top to bottom, building a prompt. The LLM only starts reasoning after it receives the resolved prompt. After the subagent finishes, the agent waits for the next utterance, which starts at start_agent again.

Example

Agent Script

1system:
2    instructions: "You are a friendly agent that helps customers with their questions."
3    messages:
4        welcome: "Hello! How can I help today?"
5        error: "Sorry, something went wrong."
6
7config:
8    developer_name: "Demo_Agent_1"
9    agent_label: "Demo Agent"
10
11access:
12    default_agent_user: "service@example.com"
13
14start_agent agent_router:
15    description: "Welcome the user and pick the right subagent"
16    reasoning:
17        actions:
18            go_to_orders: @utils.transition to @subagent.Order_Management
19                description: "Order lookups and updates"

Adapted from: Agentforce Developer Guide: Agent Script Blocks

Exam trap

welcome and error are the required system messages. A distractor may claim the config block is where you set the agent user: default_agent_user is deprecated in config and belongs in the access block.

Predict, then reveal

A customer is mid-way through an order subagent and sends a new message. Where does processing start for that message?

Show the answer

At the start_agent block (the Agent Router). Every request begins there; by default the router decides again which subagent fits.

Try it 5 min · an org with Agentforce Builder

Open Agentforce Studio from the App Launcher in an org with Agentforce Builder, and open or create an agent.
Switch the editor from Canvas to Script and find the system, config, start_agent and subagent blocks.
Result: you can point to the required welcome and error messages and to the router that every request starts from.

Practice questions

Ten exam-style questions on this lesson. Answer each one before you open it.

Question 1. A builder wants to change the greeting customers see when a conversation starts. Which block holds it?

A. config
B. system
C. access
D. variables